The AI Film Workflow No One is Talking About...Yet

So, I recently released an AI short film which utilized Seance 2.0. Reactions were very positive, so I thank you for that. And as promised, today we're going to do a full like masterclass breakdown of the film, including all the various production tips and techniques I picked up along the way. And per usual for these kinds of master classes, I'll go over the time it took and roughly block out the associated costs so you have a good idea of how much this might cost you if you take on a similar project. And we're really going to hone in on this this new workflow because I think this is a sneak peek into the future of AI film making. All right, dim the lights and let's go to the movies. So, first off, if you didn't catch Dragon Blue, don't worry about it. Uh we're actually going to be running it for review here in just a moment. And this version does differ in that I popped on Dehancer for this version. It's kind of like a film emulation plugin. Uh but if you want to see the clean version, uh that is also linked down below. On that note, grab some popcorn and crank the sound. They stole something from me and I'm going to get it back. [music] Neither will they. I'll give you one chance to [music] walk away. I'd hoped you'd say that. [music] [music] >> [music] [music] >> I [music] thought I heard some rats in the walls. [music] Thanks for the dance, boys. >> [sighs] [snorts] >> Because you stole something from me. >> I told you it wasn't for sale. >> [sighs and gasps] >> A ninja? That'd be fun. I've never fought a ninja before. But the thing with ninjas is they're mercenaries in pajamas. And you gave me a lot of money. A lot. You know who I am. >> Then all that money is yours. We'll square up after. >> I've missed you. I thought you weren't going to be here. >> Can you take me to the bank? I owe a ninja a lot of money. >> And the moral of the story is always pay off a ninja debt quickly. So off the bat, look, yeah, this is very proudly wearing its Kill Bill influence on its sleeve. I will totally cop to that. What's funny is that I this isn't even the short film I intended to make. It actually just started as a test and I ended up having a lot of fun. So I kind of just kept rolling with it and at some point I was like, "Well, this is pretty kill." But this also gave me the opportunity to test out some of Cream's multilingual capabilities and uh it definitely delivered there. We'll talk more about that in the production section which uh I guess starts now. So mainly our tools here are of course Cream 2.0 which I was using on Dreamina. Now I know as of this video access is a little bit wonky. It is coming. I do know that. But in the meantime, there is a lot that I think you'll be able to mine from this workflow. No ma no matter what video generator you're using. The other major tools I utilized for this was Nanobanana Pro, not two. It's a very important distinction. I'll go over that in a little bit. And interestingly, and maybe the biggest game changer of them all, was Claude. Yeah, Claude, specifically Co-work. Obviously, we round out with the usual grabag of like editing software and Topaz for the upscale. So, on the pre-production side now, if you haven't been playing around with Claude and well, co-work specifically, I do highly recommend you try it out. I know that this like looks a little intimidating and all like I just want I didn't write this like Claude wrote this. So, within this project, we have everything from templates for our like nano banana character designs, uh, our scene designs, um, and even our sea dance prompts are in here. This is all far too much obviously to sort of break down here. So, all of this is actually going to be in this week's newsletter, which will be released tomorrow. So, you'll you'll have access to all of this stuff. Claude also built out a full production tracker for me essentially to keep me on track. And it also created a safety filter guide. So, anytime that we would run across a keyword that would get blocked, um it was tracking that as well as whatever substitute, you know, keyword that we use as a token. so that you know when it wrote prompts later on it would substitute those words in. Massive times saver. But really the power with co-work is the fact that you can give Claude access to a folder on your, you know, desktop or computer and it'll keep everything organized for you. Like here's our character references, our various scenes. Like I don't do this. In fact, I mean, I'll show you what I normally do. What I normally do is have one folder with just a bunch of stuff thrown in it. I mean, you can judge me. It's fine. And I'm sure I'm not alone in this. Uh yeah, I mean, usually there's like there's zero organization. It's just all stuff lives here. So I quickly began to think of this clawed project as kind of like my my production office essentially like production and brainstorming. Um so you know, every once in a while I would just tell it, you know, go through and clean our our folder uh after generating some stuff. You know, obviously as I'm working here, I'm also in tandem uh working in our next section as well. Uh but one thing I did want to note, Claude does have a really impressive context window. In fact, this is all of the stuff that we got through. Uh, but you know, as you continue on, you will hit a point where it starts to get pretty stupid. Uh, so just, you know, tell it to package up a handoff prompt. Um, and I said, "Your tokens are getting lazy." Um, so it'll basically, uh, save everything out to that folder on your desktop or whatever. So, just open up a new project, point it at your folder, and then essentially it'll read the handoff prompt there, and you're back in business. So, now that we're all organized and the home office is all set up, uh, well, it's time to generate some reference images. For that task, uh, I ended up moving over to Luma's new agent board feature. Um, now granted, like these sort of agentic boards do seem to be popping up in a number of spots. Notably, Flora has a version of this where you can give it kind of a number of shots and then, uh, tell the agent, you know, give me some more coverage on these scenes. Recraft also has a version of this that I like. Uh we're actually going to be taking a look at that tomorrow. So yeah, I mean it's been a busy week, but I will say that I have not pushed any of them as hard as I pushed the Luma one. Um so yeah, this is essentially the project. Uh I kind of went ham on this. There's some extra stuff in here as well. You don't have to go quite this nuts, but I do have to say the Luma canvas agent is really smart. So, uh, initially I just began actually with just a brief um that, uh, I put in via text and, uh, just some, you know, shots of our character. Uh, ended up labeling her the girl, uh, instead of flamethrower girl cuz she didn't have her flamethrower. Um, and just by having like this text here, the board essentially knows who she is. And then I also dropped in the script so that the board would essentially have context and understanding of what we were creating here. Uh, and then from here, actually, this is just one of our nanobanana prompt templates over here. Um, we'll we'll talk about that in just a minute. So, this is just kind of a random one. So, initially, I will admit that I was kind of going with a spray and prey approach here, just generating a lot of images. Like, this is kind of like the arrival. Uh, we would end up with some stuff that, you know, is getting honing in on what I was looking for. Uh, interesting note here. By the way, there was one that it generated up that kind of has some characteristics of me in there. So, I wonder if I'm so closely associated with uh flamethrower girl at this point that Nano Banana just randomly throws me into things from there. Essentially, when I found something that I'd like, I just basically drag it over for like a collection of selects. Initially with these uh, you know, 2x two nanobanana images, what I was doing was actually taking them into Photoshop and then manually cropping them till I realized like, wait a minute, that's a robot's job. So, that made my life a lot easier. From there, it was I don't know, kind of basically the same deal. Talk to Claude about the script. uh generate up some prompt ideas about it. Um we ended up creating uh our four on characters here. Um and then yeah, just you know, we just kind of kept moving. Now again, sometimes you would just get some weird stuff like our guys hanging out here um or guys with masks off. So you know, it it basically the way that we ended up finding things was just generating by volume for the walk up confrontation scene. uh you know, a lot of stuff that clearly is not going to work in here. But again, because you know, we're generating so many images, eventually we're going to find stuff that works. And again, what's kind of cool about the Luma canvas is that we can go through on shots that we like and just kind of label them with the text tool. And again, the Luma agent knows, you know, what shot you're talking about. What is kind of interesting when you come out to this like helicopter views, you can see actually something that was intentional on my part is, uh, you know, this intro section was kind of all had a blue tint to it. Then we moved into kind of like this warm amber kind of look. Uh and then that followed into the boss's office which kind of had a red tone and then ultimately ended on that cold steel. So um yeah, little high fallutin there. Now one thing I do want to note is that Nano Banana Pro is uh I mean it's still not good with like spatial uh awareness um nor spatial consistency. So although like in the boss's office segment, you know, we got a lot of like accurate versions of the boss, but you know, the the room was kind of all over the place. So from there, it did take a number of rerolls after that to kind of hone in on the look. And then from there, I ended up pulling out enough selects that, you know, things kind of felt like roughly consistent at least. We did end up with this shot which uh I immediately gravitated towards and uh ended up saving this out because I wanted at least one kind of like hero moment in each one of the segments. We we'll talk about that when we get to the generation part. So I would say in total about 400 Nano Banana Pro images um for the project, but actually more than that because we're generating in the 2x two fashion. So really like you know four times that which does seem like a lot but you know at the same time Nano Banana is pretty cheap. So I do think that this like spray and prey approach uh does work really well. It's just a matter of like you have to be the one going through and you know picking your selects and essentially doing your own editing. Now while I would love to say that production was super smooth and there was a you know direct handoff where we did all of our images and then brought them over to Case that wasn't the case at all. every it was completely ramshackle and I was working clawed overtime in uh organizing that folder. But um you know once we had images um you know I would start rolling over to uh Dreamina and heavily relying on the omni model. This is where you can take multiple image references and then you know add your prompt in and it kind of ends up blending it all together. This was a obviously a claude written prompt. uh you know I told it what I wanted and uh you know Claude wrote this out for me in general seance is very good and within one or two generations I would pretty much get you know exactly what I was looking for now that said there it's not perfect uh but and oddly the places that actually failed were kind of interesting failed might not be the case more like a struggled um for example actually her getting out of the car this one was like took way longer than it should have for example we had this shot where she gets out not only on the wrong side of the car but essentially in the front seat uh and then on top of it this happens. We had this version which was close but not quite right. And then we had this version which uh old school body morph. I haven't seen that in a while. So that actually brought me a lot of joy. Ultimately the trick is uh here we have two drivers up front. Uh is as I've been saying for a while find the best moments from each generation and then edit them together afterwards. And that kind of leads us to the first hero moment of the short uh which is you know the first fight scene. So this took in total one two three that was a previous generation four. So it was four generations in total. Now as has been heavily documented CDS is pretty good at doing like these uh crazy action sequences. Um it doesn't necessarily always have a ton of logic to it but I mean it moves really fast. It will definitely do some wacky stuff. Um like that kick. That kick is amazing. Um, but and you know, a lot of times there isn't necessarily a ton of logic in there, like these characters just kind of keep respawning. And as a number of you had noted, uh, in the original release of this, like there is this moment where like she kind of almost like lops this dude's head off, something goes flying, but then when he comes back around, he's like, I'm fine. Um, now I probably could have gone through and rotoscoped uh that chunk of whatever um out, but you know, it just really kind of wasn't worth it. Um, the other thing, if you've got a really keen eye, you will catch some moments of like body morphing happening here. [snorts] That said, again, I think that things are happening so quickly that very rarely does anybody even notice it when it's there. So, again, I think that you the best bet here is to generate a couple of different fights, um, find the best moments, bring them into your editor, and then just kind of treat it as a bit of a puzzle to put together. Hero moment number two was, of course, the uh, silhouette fight sequence. I really wanted to do this and was initially running into like a lot of problems. Uh because I mean this is this is a tough scene. The frustrating part is that you know a lot of times we would get very close and then something like the doors would open or we would get a version like this which is I mean this is actually really good. There's even like a third dude that shows up and I I don't mind that at all but then that happens. Um and I was very committed to trying to get this as a essentially a wer. Ultimately the trick became using this frame as an input image and then using the omni model uh telling it to create a video based on 10 seconds before this happens. Uh and then that's how we ended up getting the approach of the bad guys coming in and then from there using that scene and using the omni models extend feature uh to get this scene out. Now, I I will note that I ended up prompting this in Chinese uh mostly because I was getting a lot of kickback on um you know, the blood splatter. It doesn't work all the time, but you know, sometimes that's the thing that gets you through. Now, one thing I do want to note is that there's always like this slight hiccup between extensions. Um again, I was really committed to trying to get this off as a wer. Um so, the way that I ended up kind of getting around that, and admittedly, this is kind of the best that I could get it. um is that I actually changed the frame scale here to 107. So at 100% normally you're going to see like a bit of a you'll see that shift that's there. So the number that I found was 107. I don't know if that will always be the case, but it worked for here. Um, and then I also ended up putting a slight film dissolve in here. And then uh another cross dissolve here just to kind of smooth out that like static rigidness that often happens when you're when you're kind of smashing uh an extension together. So, uh, best bet here is to actually put a cutaway in to hide it. But, uh, again, I was committed to the wer. And then finally rolling out to our third hero shot um Katana Girl in the uh frame of the boss's sunglasses here. Um this is one where I you know again going back to the multilingual thing. If you just prompt in Japanese um the character will speak in Japanese. So this this works across any of the model or any of the languages that Cance supports. So um and yeah you can get back and forth as well. If she spoke here I would just prompt her to say something and write it in English. Now, I will note that the boss office scene uh while that ended up taking uh three generations, I did I was running into like weird kickback uh again uh with like four of them where it was just like uh there's a there's an image here that uh is contains uh something that violates community standards. It was funny was that the image that it was giving me grief on was actually generated by it. This is a screenshot of uh yeah, you can see the AI up in the corner there. Um, so it actually made this and said like, "Oh, you're not allowed to use that." Look, at the end of the day, I mean, uh, Cance 2.0 is very good. It is pretty much the best video generator around right now, but it is still AI video, so it's going to do like weird things, uh, weird and mysterious things. It still, I mean, it's AI video. It has a mind of its own for full post-prouction and editing. I mean, in all honesty, pretty standard edit in all honesty. Nothing nothing really crazy in here. Uh beyond the fact that you know subtitles that was that that that one was new for me. Um but other than that yeah I mean pretty standard. Uh music was generated up in sunno. Um and other than that the only other thing that I did that was kind of like different. Uh was that I did take all of the dialogue and then ran that through Adobe Podcast Enhance uh and then brought it back in. Uh other than that you know all of the diagetic sound effects were all you know still in. Um, so yeah, it was just essentially laid directly on top. Now, I will say, of course, there are continuity errors. As many had pointed out in this sequence, the flamethrower flips around. Surprisingly, one that nobody in the comments caught was that, uh, when she walks into the compound after the fight, uh, her sword is actually sheathed. So, um, got that one past you. The point being is that you don't go through an edit like this without seeing every single mistake. So, I thank you for pointing them out. I mean, I see them, too. One that a few people had pointed out was that at the end when she gets into the cab that she's actually driving herself off. That's actually incorrect. In Japan, they drive on the other side of the road. So, I was actually really nitpicky about that one. Moving on to production time and cost. Uh, the whole project took about 2 days. Technically, three calendar days, but that's also because I'm I'm running the YouTube channel full-time. Uh, and this was just me. Well, me and Claude and the Luma agent helping out. Generation-wise, uh we came out to about 85 shots generated. Uh and then ultimately ended up using uh 32 shots. So, uh rough shooting ratio would be 2.66 out of those 85 shots. I'm not sure what like final pricing on Dreamina is going to be. That hasn't been released yet. Um so, I mean, let's just go high and ballpark it at a dollar. Um so, that would be 85 bucks. Now, I definitely went way overboard with my Nano Banana shots. I did end up using the 2K version, which is cheaper, and you know, with Nano Banana, like prices are all over the place. So, let's just use the normal API cost. So, um let's say I generated about 400 images. So, that is essentially about 50 bucks. If you are going to utilize Claude, which is was a real lifesaver in this, so I do highly recommend it. uh a Claude subscription in order to use co-work is going to cost you $17 a month. Uh definitely look into it. It's Claude is really good. So for the music, which would have been an additional $10. I did run everything through Topaz. I am actually grandfathered into the old plan, but I think the sub cost is like 25 bucks a month. Now you can also yo you bring your own upscaler. Uh and then on top of that, of course, you know, I was editing in Premiere. You can edit whatever, so I'm not going to factor that in. So total cost somewhere around $187 give or take. And look, even if we double that cost, like you know, a rough breakdown of essentially what this would have cost via Gemini. So if we were to take this at like a Hollywood level budget, Gemini estimated at around $350,000, uh, which I guess that's about right. Above the line would be 35,000. Below the line, uh, $180,000, and then, uh, you know, we had some post-production work at $70,000. um catering. Got to have catering. Uh insurance, etc. $65,000. And then I I had it run it again at sort of like the absolute indie budget level. Uh and it came up with about $35,000. So uh 3,000 for above the line, uh 18500 for uh below the line, including stunt people and whatnot. Um and the below line cost at 8,500. So yeah, I mean it sounds about right. I am also not a line producer, so you know, feel free to push back on that price if you want. Now, of course, the question is is like, is this as good a movie as one shot with professional actors, a full stunt team, and a top-notch crew? No, it isn't. But there's also a massive cost difference here. And look, a common criticism laid out is that an AI film isn't a real movie. Like, you can't put this up against a Nolan film or a Kubric movie or or even a Tarantino film. I fully agree, but also, I made this for less than $200 in 2 days. Just me. As for your enjoyment of it or its artistic merit, I mean that's a whole other thing. But no, it's it's not a Nolan movie or a Scorsesei movie or a Kurissawa movie. But if you're taking the stance that AI filmm will never replicate real film making, well, I mean, you kind of have to remember that Shark NATO 5 exists, that is a real movie that actually exists and is part of the tapestry of cinema. So, look, at the end of the day, I had fun making this. I hope you had fun watching it. Is it perfect? No, it's not. But it's a good bar of where we are right now in 2026. And I've been at this, you know, since it all kicked off about 3 years ago. And I can tell you the bar for like industry standard. I mean, it's just constantly moving. But it also doesn't matter. You know, make something you want to see. Put it out in the world. Some people are going to like it, some people won't. But the important part is that you don't have to spend 5 years pitching and and begging to make your thing anymore. You can just do it. Anyhow, I'm going to get off my soap box now. I got a rap party to get to. As always, I thank you for watching. My name IS TIM. >> THAT'S A WRAP. [screaming] >> THANK YOU GUYS SO MUCH. >> Good job. >> WE LOVE YOU GUYS. >> GOOD JOB. >> We'll see you next time. >> See you next time. Have a good night.